Dimension reduction techniques in track geometry quality analysis and safety
Date
2019
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Federal Railroad Administration (FRA) annual safety data has consistently shown that 30-40% of train accidents are track-related. The FRA data further suggests that 15% of all major accidents are due to track geometry-caused derailments. Track quality has been classically examined by the use of Track Quality Indices (TQIs) while geometry safety is monitored by FRA threshold violations. The role of Track Quality Indices in track safety needs a close re-examination because it raises the question: if TQIs are used to assess track quality (and determine geometry maintenance regimes), why are they not sensitive to safety exceptions? In other words, an index that measures the quality of track should inherently have the ability to imply when rail track is not safe for operation, hence, safety. The current state of practice suggests that geometry exceptions be handled by issuing slow-down orders or closure on a track section that violates FRA safety threshold until the exception is corrected. Is there anyway the resulting spot corrective maintenance can be addressed by TQIs that are both sensitive to track quality and safety? The aim of this dissertation is to re-evaluate the purpose of TQIs and implement novel machine learning applications to help integrate elements of track quality, geometry exceptions and safety. ☐ Track geometry data exhibits classical big data attributes: value, volume, velocity, variability, visualization, veracity and variety. Track Quality Indices-TQI are used to obtain average-based assessment of track segments and schedule track maintenance. TQI is expressed in terms of track parameters like surface, gage, cross-level, alignment etc. Each of these parameters is objectively important. However, understanding what they collectively convey for a given track segment can be challenging. Several railways, both passenger and freight have developed indices that combine different track parameters to assess overall track quality. Some of these railways have selectively recognized some parameters whilst dropping others. ☐ Using track geometry data from freight and passenger tracks as well as track acceleration data, this dissertation develops a framework to combine track geometry parameters into a low dimensional form (TQI). This simplifies the track properties without losing much variance in track geometry data. A linear implementation of this spurred the use of principal components (PCs) while the employed nonlinear dimension reduction is T-Stochastic Neighbor Embedding (t-SNE). To validate the use of PCs and t-SNE as TQIs, a three-phase approach was implemented. The first phase was to identify a suitable machine learning technique that works well with track geometry data. The second step was to train the identified machine learning technique on the track geometry data using raw geometry parameters, combined TQIs, PCs and t-SNE components as geo-defect predictors. The performance of the predictors were compared using true and false positive rates in the third phase. ☐ The results show that three PCs were better at predicting defects and revealing salient characteristics in track geometry data than combined TQIs for a passenger track. A freight track analysis shows that t-SNE had a higher predictive power than PCs, but the latter is well suited for geo-defect probability threshold development due to its linear nature. This study yields promising results that have the potential of reducing spot or corrective railroad track maintenance through the use of artificial TQIs that are both sensitive to track safety and ride quality.
Description
Keywords
Analytics, Data science, Machine learning, Railway, Safety, Derailments, Track Quality Indices, Geometry exceptions